TPA: Two-stage progressive attention segmentation framework for hepatocellular carcinoma on multi-modality MRI
Lei Gao1, Weilang Wang2, Xiangpan Meng2
1Institute for AI in Medicine, School of Artificial Intelligence, Nanjing University of Information Science and Technology, Nanjing, China.
Medical Physics
|February 2, 2024
Summary
This study introduces a two-stage progressive attention (TPA) framework to improve hepatocellular carcinoma (HCC) segmentation using multi-phase dynamic contrast-enhanced MRI. The TPA model effectively addresses inter-phase correlation challenges, achieving superior segmentation performance.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Oncology
Background:
- Dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) is vital for hepatocellular carcinoma (HCC) diagnosis and measurement.
- HCC heterogeneity leads to varied appearances across DCE-MRI phases, complicating segmentation.
- Inconsistent sizes and boundaries in DCE-MRI phases hinder inter-modality correlation and accurate segmentation.
Purpose of the Study:
- To design a multi-modality segmentation model for HCC.
- To learn meaningful inter-phase correlation from DCE-MRI data.
- To achieve improved HCC segmentation accuracy.
Main Methods:
- Proposed a two-stage progressive attention (TPA) segmentation framework for HCC.
- Utilized transformer architecture and radiologist decision-making principles.
- Developed a multi-modality attention transformer (MAT) module in the second stage to focus on size-representative features.
Main Results:
- The TPA model achieved a Dice coefficient of 0.822 (internal) and 0.772 (external) on a dataset of 386 HCC cases.
- Performance was particularly strong in a subgroup with weak inter-phase correlation (Dice 0.829 internal, 0.791 external).
- Outperformed state-of-the-art models, especially in challenging cases with weak inter-phase correlation.
Conclusions:
- The proposed TPA framework delivers optimal HCC segmentation results.
- Incorporating clinical prior knowledge into network design is practical and effective.
- The model demonstrates robustness in handling HCC heterogeneity and inter-phase inconsistencies.
